Are there papers showing “hivemind AI” or AIs that control other AIs?

There are a lot of papers on ‘federated training’ where different parts of the training are done on separate datasets on different computers. Also a lot of literature on doing both training and inference in ‘the edge’.

For instance a surveillance system could have an A.I. in the camera and one in the building and then another one in the cloud that oversees all the buildings.

Good question. But, with the way AI is currently designed, is it capable of this type of configuration without heavy backend modification of code to operate in a "hivemind" state? And how much do you want the Borg to become true? And will you welcome our Borg overlords?

Not within the capabilities or design of any current LLM (aka "AI"). So far the only things AIs can control are credulous humans.